Can AI handle receivables, bills due and cash flow forecasts?

Every Monday you want three answers: who owes me money, what do I owe this week, and how much cash will I have in the months ahead. Can AI handle that? Yes, with a split. AI can run your receivables check and your bills-due check on its own every week, because both are reading and comparing. AI can draft your cash flow forecast, but a person should approve it before anyone acts on it. AI should never be the one that moves money. That is the whole answer, and the rest of this page shows how to sort any finance task into one of those three groups.

How to sort a finance task: the decision tree

Ask two questions about the task, in this order.

1. Does a wrong answer move money? If the task sends a payment, issues a refund, or changes what you owe, a person stays in the loop. This is the "keep it human" branch. AI can prepare the work. A person presses the button.
2. Is the source data one clean export? An export is a spreadsheet file your bank, Shopify or accounting software lets you download. If the numbers come from one export that looks the same every week, AI can run the task by itself on a schedule. If the numbers are scattered across emails, PDFs and three systems, AI drafts and a person approves.

That sorts the three tasks you asked about.

• Receivables check (who owes you money, and are they late): the data is one export, and a wrong answer moves nothing. AI does it alone, every week.
• Bills due (what you owe and when): same two answers. AI does it alone, every week. It only reads and lists. It never pays.
• Cash flow forecast (how much cash you will have in 13 weeks): a wrong answer does not move money directly, but you will make hiring and ordering decisions from it. AI drafts it. You approve it.

Which finance tasks can AI do without a person checking each one?

The two that only read and compare: the receivables check and the bills-due list. A mistake in either one costs you a wrong list, not a wrong payment, and you will see it the next Monday. For a second pair of eyes, I wrote about having a second AI check the first.

How does AI check receivables against what customers actually paid?

In the older days you would open your invoice list, open your bank statement, and tick off each invoice against each deposit. Retailers pay late, pay short, and pay two invoices in one transfer, so you would also chase the differences by hand. And three hours later you would have your answer.

Today, give this to Claude (an AI assistant you type to, like a very fast bookkeeper) and you have your answer in thirty seconds. (A prompt is just the instructions you type in; you can paste this one as it is.) Attach two files first: your open invoices export and your last 30 days of bank deposits.

You are checking accounts receivable for a small food and beverage brand.
I have attached two files: a list of open invoices and a list of bank deposits.

Match each deposit to the invoice or invoices it paid. A deposit can cover
more than one invoice. Match on amount first, then on customer name, then on
reference numbers.

Give me four short lists:
1. Invoices that are paid in full.
2. Invoices that were paid short, with the amount missing.
3. Invoices that are more than 7 days past due with no matching deposit.
4. Deposits you could not match to any invoice.

Under each list, write one plain sentence on how sure you are. If you had to
guess a match, say so and show the guess. Do not invent invoices or amounts
that are not in the files.

To make it run without you, the same instructions become a scheduled job. A scheduled job is a task your computer runs by itself at a set time, say Monday at 6:00 AM, and the result lands in your inbox. Setting it up is a one-time job for your bookkeeper or ops person, or for us (see the end of this page). Until then, attach the two files yourself for the first few Mondays and read the four lists. After the setup, nobody attaches anything. For a seven-figure brand, that is the receivables meeting you stop having.

What does a cash flow forecast from AI need as input to be trusted?

Three things, and you supply them once.

• Your current bank balance, taken from the same export every week.
• Your open invoices with due dates, and your bills with due dates. The two checks above already produce these.
• Your known recurring costs: payroll dates, rent, loan payments, software.

With only those, the forecast is a plain adding and subtracting exercise, and AI is good at that. It gets unreliable when it has to guess, so the instructions tell it not to.

Here is an illustrative example. A made-up brand starts the week with $120,000 in the bank. It expects $85,000 in retailer payments over six weeks and owes $70,000 in bills over the same stretch. The forecast shows $135,000 at week six, with a dip in week three when payroll lands before the big retailer pays. That dip is the thing you wanted to see. (All figures here are illustrative, not real.)

The instructions for the draft:

Build a 13-week cash flow forecast for my company from the files attached:
current bank balance, open invoices with due dates, bills with due dates,
and a list of recurring costs.

Show one row per week with: starting cash, money coming in, money going out,
ending cash.
Assume a customer pays 10 days after the due date unless the file shows how
they usually pay.
Mark every number that came from an assumption and not from a file.
Flag any week where ending cash falls below $20,000.
Do not fill in any figure you were not given. Ask me for it instead.

Where should a human approve before anything moves money?

At three points, and they are the same every time.

• Before a payment is sent. AI can list the bills due this week and rank them by urgency. You or your bookkeeper approves each payment.
• Before a forecast is used to hire, order inventory or take on debt. Read the marked assumptions first. Those are where it is wrong, if it is wrong.
• Before anyone is told a customer is late. A late flag means look into it. It is not proof.

Treat the setup like a new bookkeeper. I wrote about managing AI like a hire: give it a narrow job, check the first few weeks, and widen it only when it earns that. And if the numbers from two systems disagree before AI ever sees them, that is a different problem, which I covered in which number is right.

If you would rather have this set up and running than build it yourself, that is what the Reporting Sprint does: we set up your weekly numbers so they run without you.